Visual Analysis of Magnetic Resonance Spectroscopy Imaging Data for the Study of Human Brain Tumors

(1)

Eurographics Conference on Visualization (EuroVis), Posters Track (2017) A. Puig Puig and T. Isenberg (Editors)

Visual Analysis of Magnetic Resonance Spectroscopy Imaging Data for the Study of Human Brain Tumors

M. Jawad and L. Linsen

Westfälische Wilhelms-Universität Münster, Germany

Abstract

Magnetic Resonance Spectroscopy Imaging (MRSI) is an in-vivo method for measuring metabolite concentration in various tissues. Typically, individual metabolites are examined in detail. We provide an interactive visualization tool that allows for the simultaneous analysis of all metabolite concentrations. The multi-dimensional data visualization is based on star coordinates interaction in a projected view. We derive a segmentation of the accompanying magnetic resonance image (MRI) to investigate the metabolite distribution within tissues including partial volume effect handling. Coordinated views between spatial slice- based visualizations and multidimensional metabolite spaces allow for the selection of individual voxels, certain anatomical regions, or groups of voxels with similar concentrations for a comparative analysis in the linked views as well as in further linked statistical plots. We apply our method to the analysis of brain tumors and surrounding tissue.

1. Introduction

In 2012, WHO reported 256,213 brain cancer cases from which 189,382 deaths are recorded [NIH15]. Magnetic Resonance Spec- troscopy Imaging (MRSI) allows for an improved diagnosis of the type of tumor and the infiltration of surrounding tissues that is otherwise hard to examine [BTBJ04]. While magnetic resonance imaging (MRI) provides location, shape, and size of a tumor, MRSI allows for a better classification into the more than 100 tumor types and a rating between most malignant to benign.

MRSI is a non-invasive bio-medical technique used by radiologist and neurosurgeons for in-vivo brain studies to quantitatively analyze the concentrations of affecting bio-chemicals (called metabolites) [VDMCW94]. While most analyses of MRSI data focus on individual metabolites, we present an approach to visually analyze the entire multi-dimensional metabolite space in conjunction with the imaging space, thus, exploiting the full potential of the imaging method.

2. Related Work

In clinical practice, tools like jMRUI [SDCA^∗09] are used that includes various pre-processing algorithms for MRSI data and a slice-based visualization, where for some selected metabolites the concentration of a metabolite is color-coded and the color map is overlaid with the MR image. Commercial tools like SyngoMR and SpectroView that are distributed with scanners provide similar functionality. Kinoshita et al. [KY97] showed in in-vitro experi- ments on the human brain that the list of metabolites are active in certain type of tumors goes beyond what is provided by standard tools. Feng et al. [FKLTI10] and Nunes et al. [NRS^∗14] proposed

first approached to use multidimensional data visualization methods such as scatter plots and parallel coordinates for visualizing metabolite concentrations. We build upon such methods and ex- tend in various directions including coupling it with segmentation results, visualizing the multidimensional feature space with projections which allows for a selection of voxels with similar metabolic concentrations in the multi-dimensional space, and comparative visual analysis of individual voxels and/or groups of interest.

3. Visual Encodings and Interaction Mechanisms

The spatial visualizations are based on the MR images with overlaid color maps or glyphs. We support an automatic segmentation algorithm (MICO) [LGD14] for segmenting the MR image, but a manual segmentation is also possible by simply marking respective regions in the slice. As the MR image has a substantially higher resolution than the MRS image, the MRSI voxels exhibit a partial volume effect when compared to the MRI segmentation. We vi- sualize the segmentation result at an MRSI-voxel resolution using pixel-map glyphs, where each MRSI voxel is visualized as a rect- angle that is filled with colored bars according to the segmentation result of the covered MRI voxels. As each segment is assigned one distinct color, each glyph consists of respective portions of those colors, see Figure1a.

We pre-process the MRSI data with Tarquin [RWPA06] to ex- tract all metabolite concentrations. The multidimensional space is visualized using linear projections, which can be modified using a star coordinate widgets, see Figures1band1d. Layouts by automatic projection methods such as Principal Component Analysis are also supported. Each MRSI voxel is represented by one circular

c 2017 The Author(s)

Eurographics Proceedings c2017 The Eurographics Association.

DOI: 10.2312/eurp.20171178

(2)

M. Jawad & L. Linsen / Visual Analysis of Magnetic Resonance Spectroscopy Imaging Data for the Study of Human Brain Tumors

(a) (b)

(c) (d)

Figure 1:(a) Automatic MRI segmentation result impose on MRSI voxels using pixel-map glyphs for encoding partial volume effect. (b) Linear projection of multidimensional metabolite space with interactive separation of clusters by control points (black boxes). Voxels are encoded by pie-chart glyphs of segmentation result. (c) Box plots of all metabolite concentrations for one reference group (grey matter) when compared to the tumor voxels (here average) shown as red dots. (d) Star coordinate widget showing long axes of dominant metabolites.

glyph, where the glyph is a pie chart that shows the same colors and proportions as the glyphs in the spatial slice-based view.

When investigating the multidimensional data projection, one can observe the different colors reflecting different tissue types.

One can also impose a hard segmentation, where each voxel is assigned to the segment that contributes most. A respective classification defines groups of voxels. In the projected view, the centroid of each view is computed and the classes can be separated (if possible, at all, with a linear projection) by pulling apart those control points following the approach by Mochanov and Linsen [ML14]. Voxels or regions of interest can be selected both in the slice-based view and in the projected view for further comparative investigations.

Further investigations can be performed with additional linked statistical plots such as scatterplots for a pair of selected metabolites and, in particular, by a bar chart plot over all present metabolites, see Figure1c. Comparisons between two selections are possible by selecting a reference group, which is depicted using bar charts for its metabolite concentrations, and a test object like a voxel or a group of voxels, which can be depicted by individual dots (or also bar charts).

4. Brain Tumor Study Results

We applied our visual analysis tool to data from a ¹H MRSI scan (3T Siemens scanner, TR/TE/flip = 1700ms/135ms/90) of a 26-year-old male patient with a brain tumor. (Data courtesy of Miriam Bopp and Christopher Nimsky, Universitätsklinikum,

Marburg, Germany.) The MRI volume is 224×256×144mm³ with 1 mm slice thickness and the two MRSI series each have a 160×160×12mm³ FoV (Field of View). Each MRSI voxel stretches over 10×10×12 MRI voxels. Figure1ashows the automatic segmentation result at the MRSI-voxel level with pixel-map glyphs encoding the partial volume effect. The segmentation result is not perfect and would require manual fine-tuning, but the tumor is clearly visible. Figures1band1dshow the respective projection and star coordinate configuration of the multi-dimensional metabolite space with pie-chart glyphs. The segments have been in- teractively separated. The long star coordinate axes indicate, which metabolites are mainly responsible for the separation. The typical suspects show up plus some other metabolites. Figure1cshows a comparison of the tumor voxels (represented by their mean) against grey matter to investigate for which metabolites the concentrations differ. This indicates what type of tumor this may be. In a next step, one would select neighboring voxels to the tumor to investigate their concentrations and judge whether they are already partially infiltrated by the tumor.

5. Conclusions

We have presented a novel tool for the investigation of all metabolite concentrations measured by MRSI to exploit its full capacity.

Visual encodings in spatial dimensions and metabolic space al- lowed for coordinated interaction and comparative visual analyses.

The tool showed potential in supporting the analysis of tumor types and voxels surrounding the tumor area.

c

2017 The Author(s) Eurographics Proceedings c2017 The Eurographics Association.

98

(3)

M. Jawad & L. Linsen / Visual Analysis of Magnetic Resonance Spectroscopy Imaging Data for the Study of Human Brain Tumors

References

[BTBJ04] BURNETN. G., THOMASS. J., BURTONK. E., JEFFERIES S. J.: Defining the tumour and target volumes for radiotherapy.Cancer Imaging 4, 2 (2004), 153.1

[FKLTI10] FENGD., KWOCKL., LEEY., TAYLORII R. M.: Linked exploratory visualizations for uncertain mr spectroscopy data. In IS&T/SPIE Electronic Imaging(2010), International Society for Optics and Photonics, pp. 753004–753004.1

[KY97] KINOSHITA Y., YOKOTA A.: Absolute concentrations of metabolites in the human brain tumors using in vitro proton magnetic resonance spectroscopy.NMR in Biomedicine 10, 1 (1997), 2–12.1 [LGD14] LIC., GOREJ. C., DAVATZIKOSC.: Multiplicative intrinsic

component optimization (mico) for mri bias field estimation and tissue segmentation.Magnetic resonance imaging 32, 7 (2014), 913–923.1 [ML14] MOLCHANOVV., LINSENL.: Interactive Design of Multidi-

mensional Data Projection Layout. InEuroVis - Short Papers(2014), Elmqvist N., Hlawitschka M., Kennedy J., (Eds.), The Eurographics As- sociation.doi:10.2312/eurovisshort.20141152.2 [NIH15] NIH: Adult Central Nervous System Tumors Treatment for

Health Professionals.Natinal Cancer Institute(2015). Web Notes.1 [NRS^∗14] NUNES M., ROWLAND B., SCHLACHTER M., KEN S.,

MATKOVICK., LAPRIEA., BUHLERK.: An integrated visual analysis system for fusing mr spectroscopy and multi-modal radiology imaging.

InVisual Analytics Science and Technology (VAST), 2014 IEEE Confer- ence on(2014), IEEE, pp. 53–62.1

[RWPA06] REYNOLDSG., WILSONM., PEETA., ARVANITIST. N.:

An algorithm for the automated quantitation of metabolites in in vitro nmr signals.Magnetic resonance in medicine 56, 6 (2006), 1211–1219.

1

[SDCA^∗09] STEFAND., DICESAREF., ANDRASESCUA., POPA E., LAZARIEVA., VESCOVOE., STRBAKO., WILLIAMSS., STARCUK Z., CABANASM.,ET AL.: Quantitation of magnetic resonance spectroscopy signals: the jmrui software package.Measurement Science and Technology 20, 10 (2009), 104035.1

[VDMCW94] VION-DURYJ., MEYERHOFFD., COZZONEP., WEINER M.: What might be the impact on neurology of the analysis of brain metabolism by in vivo magnetic resonance spectroscopy? Journal of neurology 241, 6 (1994), 354–371.1

c 2017 The Author(s)

Eurographics Proceedings c2017 The Eurographics Association.

99